AI Video Glossary: Every Term Explained

We maintain a comprehensive glossary at apostle.io/learn. This post walks through the most important terms with production context — what they actually mean when you are building AI video for clients.

Generation Techniques

These are the foundational methods for creating AI video. Every project we deliver uses at least one of them, and most use several in combination.

Text to Video

Text to video is the generation method most people think of first: you write a natural language prompt and the model produces a video clip. In practice, this means writing detailed descriptions of scene composition, camera movement, lighting, and subject action, then iterating on the prompt until the output matches the creative brief. We use text-to-video primarily for concepting and for abstract or environmental shots where there is no specific reference to match.

Image to Video

Image to video is the single most common workflow in professional AI video production. You provide a reference image — a photograph, a design comp, a product render — and the model animates it into motion. In practice, this means we can take a client’s existing brand photography and bring it to life with cinematic camera movement and subtle environmental animation. The reference image gives the model far more visual information than a text prompt alone, which is why the results are consistently more controlled and on-brief.

Video to Video

Video to video takes existing footage and transforms it through a generative model. This could mean changing the visual style of live-action footage, converting rough animatics into polished sequences, or applying a completely new aesthetic to documentary material. In practice, this is powerful for clients who have existing footage they want to reimagine without reshooting. The source video provides both the motion data and the compositional structure, so the output maintains natural movement while adopting an entirely new look.

Inpainting

Inpainting replaces a selected region within a frame while keeping everything else intact. You mask the area you want to change — an unwanted object, a logo, a background element — and the model generates new content that blends seamlessly with the surrounding pixels. In practice, this means we can remove distracting elements from generated footage, swap out product variants in an existing scene, or fix visual artefacts without regenerating the entire clip.

Outpainting

Outpainting extends a video frame beyond its original boundaries. If you have a tightly cropped shot that needs to become a wide shot, outpainting generates the missing visual information around the edges. In practice, this is most useful when adapting aspect ratios — taking a 16:9 horizontal video and extending it to 9:16 vertical for social platforms without simply cropping and losing the core composition.

Style Transfer

Style transfer applies the visual characteristics of one image or video to another. You provide a style reference — a painting, a film still, a mood board image — and the model reinterprets your source content in that aesthetic. In practice, this means a client can say “we want this to feel like a Wes Anderson film” or “match the colour palette of our brand guidelines,” and we can apply that visual language systematically across an entire video campaign.

Motion Transfer

Motion transfer extracts the movement from one video and applies it to a different subject or scene. A dancer’s performance can drive the motion of an animated character. A hand gesture captured on a phone can control a product animation. In practice, this gives us a way to direct AI-generated subjects with the specificity of a live-action shoot — we capture the performance we want, then transfer it onto the visual we need.

Camera and Movement

AI video models have become remarkably capable at simulating real cinematography. These terms describe the virtual camera controls that give directors precision over how a scene is captured.

Camera Control

Camera control refers to the ability to specify how the virtual camera moves during generation. This includes direction, speed, trajectory, and focal behaviour. Runway Gen-4 excels here, offering granular control over camera paths that was simply not possible in earlier models. In practice, camera control is what separates footage that looks “AI-generated” from footage that looks deliberately directed.

Dolly Zoom

The dolly zoom — sometimes called a vertigo shot — simultaneously moves the camera toward or away from a subject while adjusting the focal length in the opposite direction. This creates the distinctive effect where the subject stays the same size but the background appears to stretch or compress. In practice, AI models can now replicate this classic technique through prompt direction or camera path specification, adding dramatic emphasis to key moments.

Tracking Shot

A tracking shot follows a subject as they move through a scene, keeping them in frame while the background shifts. In AI video, this requires the model to maintain both the subject’s appearance and the spatial logic of the environment over multiple seconds. In practice, tracking shots are one of the best tests of a model’s temporal consistency — if the model cannot hold the scene together across frames, a tracking shot will expose it immediately.

Crane Shot

A crane shot moves the camera vertically, typically starting low and rising to reveal a wider scene, or descending from an elevated position into the action. AI models handle crane shots well because the gradual perspective shift gives the model time to build spatial coherence. In practice, we use AI crane shots frequently for establishing sequences and reveal moments in brand films.

Pan and Tilt

Pan refers to horizontal camera rotation; tilt refers to vertical rotation. These are the simplest camera movements, and AI models execute them reliably. In practice, pans and tilts are the workhorses of AI video production — they add cinematic dynamism to a scene without the complexity of full camera translation, which means fewer opportunities for visual artefacts.

Handheld Camera Effect

The handheld camera effect adds the subtle, organic instability of a human-operated camera. Slight drifts, micro-corrections, and natural wobble that make footage feel immediate and documentary-like. In practice, this is useful for social content, testimonial-style videos, and any brief where the client wants authenticity over polish. Most models can approximate this through prompt direction.

Aerial Shot

An aerial shot captures the scene from above, simulating a drone or helicopter perspective. AI models are particularly strong here because aerial footage involves broad, sweeping landscapes with less fine detail to maintain at close range. In practice, AI aerial shots can replace expensive drone permits and flight logistics, delivering establishing shots and transitions that would otherwise add thousands to a budget.

Quality and Post-Production

These terms relate to the technical quality of generated footage and the processes for refining it after initial generation.

Temporal Consistency

Temporal consistency is the degree to which visual elements maintain their appearance, position, and physics across consecutive frames. When temporal consistency breaks down, you see the characteristic “morphing” or “flickering” that marks AI footage — a face subtly shifts shape, a building wobbles, textures swim. In practice, temporal consistency is the single most important quality metric for professional AI video. It is the difference between footage a client can use and footage that looks obviously synthetic.

Video Upscaling

Video upscaling increases the resolution of generated footage, typically from 1080p to 4K. AI upscalers do not simply enlarge pixels — they use trained models to intelligently add detail and sharpness. In practice, most AI video models generate at 720p or 1080p natively. Upscaling is a standard post-production step that brings the output to broadcast or digital cinema resolution requirements.

Frame Interpolation

Frame interpolation generates new intermediate frames between existing ones, increasing the frame rate for smoother motion. If a model outputs at 12 frames per second, interpolation can bring it to 24 or even 60 fps. In practice, this is essential for footage that will be used in professional contexts — clients expect smooth, broadcast-quality motion, and native AI output frame rates are often not sufficient.

Lip Sync

Lip sync matches mouth movement to spoken audio, generating realistic articulation that aligns with dialogue. Veo 3.1 has set the benchmark here, producing lip-synced speech that holds up in close-up shots. In practice, lip sync capability determines whether you can use AI-generated human characters for dialogue-driven content. Without it, you are limited to voice-over narration or non-speaking roles.

Character Consistency

Character consistency is the ability to maintain a character’s exact appearance — face, body, clothing, proportions — across multiple generations. If you generate ten shots of a brand mascot, character consistency determines whether they look like the same entity in every frame. In practice, this remains one of the hardest problems in AI video. We use reference images, LoRA models, and multi-pass generation workflows to maintain consistency across longer-form content.

Super Resolution

Super resolution is the broader term for AI-driven resolution enhancement. While closely related to upscaling, super resolution can also recover detail from degraded or compressed source material. In practice, this is useful when working with client-supplied assets that were not captured at high resolution — old footage, screen recordings, or compressed web downloads that need to meet broadcast quality standards.

Audio

AI video production does not stop at the visual. Audio generation and synchronisation have become integral to delivering complete, client-ready content.

Voice Cloning

Voice cloning creates a synthetic reproduction of a specific person’s voice from a sample recording. The cloned voice can then speak any script with the original speaker’s tone, cadence, and characteristics. In practice, this enables localisation at scale — a single voice-over performance can be cloned and regenerated in multiple languages while maintaining the speaker’s identity. We do not use voice cloning without explicit consent and contractual authorisation from the original speaker.

AI Sound Design

AI sound design generates sound effects, ambient audio, and atmospheric textures using generative models. Rather than pulling from stock libraries, the model creates original audio matched to the visual content. In practice, this means generating the exact environmental sound a scene requires — a specific quality of wind, a particular room tone, mechanical sounds that match on-screen action — without licensing costs or library limitations.

Audio Synchronization

Audio synchronisation aligns generated or recorded audio precisely with visual events in the video. Footsteps land on the right frame. Music accents hit on scene transitions. Sound effects match on-screen impacts. In practice, synchronisation quality is what makes the difference between AI video that feels like a rough cut and AI video that feels finished. Automated sync tools have improved dramatically, but manual adjustment is still part of our post-production workflow.

Model Architecture

For the technically curious, these are the underlying systems that make AI video generation possible.

Diffusion Model

A diffusion model is the architecture behind most current AI video generators. It works by learning to reverse a noise-addition process: during training, the model learns how structured video degrades into random noise, then during generation, it starts from noise and progressively refines it into coherent footage. In practice, understanding diffusion matters because it explains why generation takes time (each frame requires many refinement steps) and why certain artefacts occur (the model occasionally fails to fully resolve noise in complex areas).

LoRA

LoRA (Low-Rank Adaptation) is a method for fine-tuning a large model on specific visual concepts without retraining the entire network. You can train a LoRA on a character’s face, a brand’s visual style, or a product’s appearance using relatively few reference images. In practice, LoRAs are how we solve character consistency for recurring subjects. We train a LoRA on the client’s product or brand character and use it across every generation to maintain visual identity.

ControlNet

ControlNet adds structured control inputs to a diffusion model — depth maps, edge detection, pose estimation, segmentation maps. These inputs guide the generation process without overriding the model’s creative capabilities. In practice, ControlNet is how we go from “generate something roughly like this” to “generate exactly this composition with this pose and this depth structure.” It bridges the gap between creative AI and directed production.

Latent Space

Latent space is the compressed mathematical representation where diffusion models actually operate. Rather than working directly with pixels, the model encodes visual information into a much smaller latent representation, performs the generation process there, and then decodes back to pixel space. In practice, this is why AI video generation is even feasible — working in latent space reduces the computational load by orders of magnitude compared to operating directly in pixel space.

Transformer Architecture

The transformer is the neural network architecture — originally developed for language processing — that now underpins the most capable video generation models. Transformers process sequences of data using attention mechanisms that can relate any element to any other element, regardless of distance. In practice, transformer-based video models handle long-range temporal relationships better than older architectures, which is why recent models maintain coherence over longer clip durations.

Production Workflow Terms

These are the concepts that come up every day in our production pipeline. If you are working with an AI video studio, you will encounter all of them.

Prompt Engineering for Video

Video prompts differ significantly from image prompts. An image prompt describes a single moment; a video prompt must describe motion, temporal progression, camera behaviour, and scene dynamics across multiple seconds. In practice, effective video prompting is a skill we have developed through thousands of generations. The prompt structure, level of detail, and sequencing of information all affect output quality in ways that are not immediately obvious.

Reference Image

A reference image is the starting visual input for an image-to-video generation. It establishes the composition, colour palette, subject appearance, and environmental context that the model will animate. In practice, the quality and preparation of the reference image is the single largest determinant of output quality. We often spend more time preparing reference images than we do on the generation itself — colour grading, compositing, and art directing the reference to ensure the model has the best possible starting point.

First Frame / Last Frame

Some models accept both a first frame and a last frame, generating the video transition between the two states. This gives you control over both where the shot begins and where it ends. In practice, first-frame/last-frame control is extremely valuable for narrative work — you can define the emotional arc of a shot by specifying its start and end states, and the model interpolates the journey between them.

Batch Rendering

Batch rendering is the process of queuing multiple generations to run sequentially or in parallel, producing large volumes of output without manual intervention. In practice, we use batch rendering when a project requires dozens or hundreds of variations — different product colours, multiple scene options for client selection, or systematic exploration of prompt variations to find the optimal output.

Seed Control

The seed is a numerical value that initialises the random noise pattern used at the start of generation. Using the same seed with the same prompt and settings produces identical output. In practice, seed control is essential for reproducibility. When we find a generation we like, we record the seed so we can reproduce it exactly, make incremental prompt adjustments while keeping other variables constant, or regenerate at higher quality settings.

CFG Scale

CFG (Classifier-Free Guidance) scale controls how strictly the model follows the prompt versus how much creative latitude it takes. A high CFG scale produces output that adheres closely to the prompt but can look rigid or over-saturated. A low CFG scale produces more natural, varied results but may drift from the brief. In practice, finding the right CFG scale is part of every generation session — we typically start at a moderate value and adjust based on whether the output needs more fidelity or more organic quality.

Business and Licensing

The commercial and legal terms that determine what you can actually do with AI-generated content in a professional context.

Commercial License

A commercial licence grants the right to use AI-generated content in revenue-generating contexts — advertising, branded content, product marketing, broadcast media. Not all AI tools grant commercial rights by default; some restrict generated content to personal or non-commercial use. In practice, commercial licensing is the first thing we verify before onboarding any new tool into our production pipeline. If the licence does not explicitly permit commercial use, the tool does not enter our workflow regardless of its capabilities.

Credits System

Most AI video platforms charge through a credits system rather than flat subscription fees. Each generation consumes a number of credits based on resolution, duration, and model complexity. In practice, the credits system means that production budgeting for AI video requires careful calculation — we model credit consumption per shot, per scene, and per project to provide accurate cost estimates before production begins.

AI Video Cost Per Second

Cost per second is the metric that translates abstract credit costs into a unit clients can understand and compare. It accounts for credits consumed, subscription costs, failed generations, and iteration cycles to produce a true cost figure per second of usable output. In practice, we track cost per second across every tool we use, updated monthly, because pricing and model efficiency change frequently. This metric drives our tool selection for each project.

Content Policy

Content policy defines what each AI platform permits and prohibits in generated content. This includes restrictions on violence, sexual content, real likenesses, political content, and brand impersonation. In practice, content policy matters because it can block an entire creative direction mid-production. We review content policies before committing to a tool for a specific project to ensure the creative brief is achievable within the platform’s restrictions.

Terms That Matter Most in Production

Of everything above, five terms come up more than any others in real client conversations. These are the concepts that directly affect whether a project succeeds.

Temporal Consistency is what clients notice first when something is wrong. If the footage morphs, flickers, or wobbles, nothing else matters — the technical quality overrides the creative quality. Every tool selection and workflow decision we make optimises for temporal consistency above almost everything else.

Character Consistency is the barrier to longer-form AI content. A single shot can look flawless, but maintaining that same character across 30 shots in a brand film requires deliberate technique — LoRA training, reference image discipline, and multi-pass quality control. When clients ask “can you make a character appear throughout the whole video,” this is the capability they are asking about.

Camera Control is what transforms AI video from a novelty into a production tool. The ability to specify a dolly zoom, a tracking shot, or a crane movement means a director can actually direct. Runway Gen-4 pushed this forward significantly, and every major model now competes on controllability.

AI Video Cost Per Second is the number that makes AI video production viable as a business model. When traditional production costs $1,000 to $5,000 per finished second and AI-native production costs a fraction of that, the economic argument writes itself. But only if you track the real cost including iteration, failed generations, and post-production, not just the raw credit price.

Commercial License is the term that determines whether any of the above matters at all. If you cannot legally use the output in client work, the technical capability is irrelevant. We have seen teams build entire campaigns on tools that did not grant commercial rights, only to discover the problem at delivery. Check the licence first.

The complete glossary covers all 97 terms in depth. If you are evaluating AI video for your brand or building a production workflow, it is the most thorough reference we have published. And if you would rather skip the reading and talk to practitioners directly, get in touch.